Richard Sandiford writes:

Richard Sandiford writes:
However, retesting on m68k showed up a problem that had appeared in
uClibc since the last time I tried.  Specifically, revision 15785 did:

  -#define HEAP_GRANULARITY       (sizeof (HEAP_GRANULARITY_TYPE))
  +#define HEAP_GRANULARITY       (__alignof__ (HEAP_GRANULARITY_TYPE))

  -#define MALLOC_ALIGNMENT       (sizeof (double))
  +#define MALLOC_ALIGNMENT       (__alignof__ (double))

The problem is that

  (a) MALLOC_HEADER_SIZE == MALLOC_ALIGNMENT
  (b) the header contains a size value of type size_t
  (c) sizeof (size_t) is 4 on m68k, but...
  (d) __alignof__ (double) is only 2 (the largest alignment used on m68k)

So we only allocate 2 bytes for the 4-byte header, and the least
significant 2 bytes of the size are in the user's area rather than
the header.  The patch below fixes that problem by redefining
MALLOC_HEADER_SIZE to:

  MAX (MALLOC_ALIGNMENT, sizeof (size_t))

(but without the help of the MAX macro ;)).  However, we really would
like to have word alignment on Coldfire.  It makes a big performance
difference, and because we have to allocate a 4-byte header anyway,
what wastage there is will be confined to the end of the allocated block.
Any wastage will also be limited to 2 bytes per allocation compared to
the current alignment.

I've therefore used the __aligned__ type attribute to create a double
type that has at least sizeof (size_t) bytes of alignment.  I've
introduced a new __attribute_aligned__ macro for this.  It might seem
silly protecting against old or non-GNU compilers here, but the extra
alignment is only an optimisation, and having the macro is more in the
spirit of the other attribute code.